Filtering and Estimation with Antagonist Hypothesis for Failure Detection

This paper is concerned with the problem of a failure detection for one dimensional systems evolving in a steady state. The detected failures are characterized by a second order model. To defect a failure wecompute the innovation sequences generated by two filters elaborated with the two hypothesis: - H0 normal mode - H1 failure mode Under H0 hypothesis we suppose that process is corrupted by a white gaussian noise. The model of the process is known and we calculate the innovation sequence generated by a kalman filter. Under H1 hypothesis we suppose that a failure exist and we estimate the parameters of the failure model in order to compute the innovation sequence. To detect the presence of a failure, a test on the two innovation sequences is proposed. A pratical application concerning the detection of surface defect of steel products illustrates the proposed method.